/SeSe-Net

SeSe-Net: Self-Supervised Deep Learning for Segmentation

Primary LanguagePython

#SeSe_Net: semi-supervised segmentation by deep learning

#by Zeng Zeng, Yang Xulei, Yu Qiyun, Yao Meng, Zhang Le,

#Pattern Recognition Letter, Aug. 2019

To repeat the results in the paper:

  1. Download one of the three datasets used in the paper, e.g., carvana dataset

  2. Create the following folder structures

    ../input

    ../input/car_data/

    ../input/car_data/train_images

    ../input/car_data/train_masks

    ../input/car_data/train_list.csv

  3. Run sse_train_step1.py (labelled samples)

    3-1) train a unet model to generate various masks

    3-2) train a resnet model by using the generated masks

  4. Run sse_train_step2.py (un-labelled samples)

    4-1) resnet predict un-labelled samples and generate loss values to refine unet model